Who Can You Hide Behind? Author-Conditioned Style Transfer and the Evasion of a Commercial AI-Text Detector
arXiv cs.CL · 16 Nov 2026, 03:59 AEST
A study examined the effectiveness of author-conditioned style transfer in evading AI-text detectors. The results showed significant reductions in detection scores for rewritten AI-generated text.
Key facts
- AI-text detectors can be sensitive to changes in writing style.
- We examined this sensitivity by rewriting 991 AI-generated answers with Gemini, using prompts that included a writing sample from one of 55 named authors.
- The median score fell from 100% to 0%, and the mean fell from 67.9% to 11.2%.
- Of the rewrites, 66% received a score of exactly 0%.
- Scores varied across the authors used in the prompts.
Read aloud
A study examined the effectiveness of author-conditioned style transfer in evading A I-text detectors. The results showed significant reductions in detection scores for rewritten A I-generated text. A I-text detectors can be sensitive to changes in writing style. We examined this sensitivity by rewriting nine hundred ninety one A I-generated answers with Gemini, using prompts that included a writing sample from one of fifty five named authors. The median score fell from one hundred percent to zero percent, and the mean fell from sixty seven point nine percent to eleven point two percent. Of the rewrites, sixty six percent received a score of exactly zero percent. Scores varied across the authors used in the prompts.
Sources in this cluster
- Who Can You Hide Behind? Author-Conditioned Style Transfer and the Evasion of a Commercial AI-Text Detector
arXiv cs.CL · 16 Nov 2026, 03:59 AEST